## May I Have Some Neural Networks with My Insurance Data, Please?

Machine learning techniques, particularly Artificial Neural Networks (ANNs), have enjoyed an upsurge in popularity and practical applications in a myriad of disciplines. The explosion in the variety and volume of available data, coupled with cheap data storage and fast computing power, have placed ANNs front and center in data scientists’ tool boxes.

The array of ANNs’ applications is astonishingly broad: autonomous vehicle driving, speech and text recognition, image processing, medical diagnosis, financial markets and weather forecasting, to name just a few. A great deal of research and development over the past 10-20 years has rekindled the excitement surrounding ANNs, which were born from a 1940s concept designed to digitally replicate the biological mechanism of the human brain and its neurons.

But while ANNs have been on the upswing in a variety of fields, the insurance sector has not utilized these “brain-like” techniques on a large scale. Over the years, experts in insurance analytics have bolstered research and experimentation efforts with ANNs; notwithstanding, the insurance field is still heavily skewed in favor of the more familiar and traditional data mining techniques such as Generalized Linear Models (GLMs) and rule-based algorithms.

Unquestionably, the stumbling block for the adoption of ANNs in insurance has its roots in their apparent lack of interpretability – they have been defamed as “BLACK BOXES,” performing mysterious wizardry with numbers and unable to reveal their magic with easily explainable formulas and rules. At the same time, their utility and strength in modeling complex multivariate relationships have been widely recognized and praised. It appears ANNs have been reduced to an oxymoronic concept that remains out of actuarial reach due to its enigmatic nature.

But isn’t it time we finally crack open the “BOX” to take advantage of ANNs’ powerful predictive abilities? In many insurance applications, including retention modeling, fraud detection, claims triage, even traditional pricing models, researchers have shown that ANNs are useful tools that can outperform or at least enhance the effectiveness of their more ubiquitous counterparts like GLMs.

Using the “brain-like” analogy to associate the mathematical mechanism of ANNs with the complex physiology of the brain does not make the case for their use any easier. Indeed, ANNs were initially conceptualized to mimic neurons’ behavior in the way they process data. But we can simply think of an ANN (the supervised kind) as a collection of nested regression models residing in its layers. The “neurons” in the first layer represent the input (predictor) variables, which are used to build the first set of regression models. The predictions from these regressions are in turn used as inputs into the next series of regressions in the following (hidden) layers of “neurons,” and so on. When the “signal” reaches the final (output) layer of the ANN, another regression is applied to produce the final predictions. Thinking of an ANN as a set of regression models, instead of some sort of “brain-like” model, at least psychologically brings us a little closer to its older cousins, the GLMs.

Since GLMs are based on a single regression equation, it is said that their predictions are more readily explainable via the regression coefficients. ANNs, on the other hand, as collections of nested regressions, are not as easily amenable to such interpretations. It could be argued, however, that since the variables we use in our insurance models are often highly correlated, and our GLMs never exactly represent the true data generation process, GLM parameters are really not as interpretable as they are touted to be.

Because of ANNs’ “black-boxiness,” various techniques have been proposed to unpack their output so it can be more easily compared with that of other models, including sensitivity analysis, NID (Neural Interpretation Diagram), Fuzzy Logic theory for rule extraction, or even “modeling” the ANN’s predictions to derive a simple mathematical expression. These simplifications can be used to interpret the effect and relative importance of the explanatory variables. These approaches, together with the regression-like interpretation of ANNs, should remove some of the mystery from these models.

Traditional predictive modeling techniques like GLMs are here to stay due to their wide range of applicability and relative ease of interpretation. But more advanced data mining techniques like ANNs should not be shunned, but rather embraced as viable contenders by responsible modelers in their efforts to build effective predictive models. Perhaps instead of treating these different techniques as competing models, we can instead apply ensemble methods such as “stacking,” in which the predictions of several different models are combined to capture the best of all worlds.

Now, I believe I ordered some neural networks with my data….

*Radost Wenman, FCAS, MAAA, is a Consulting Actuary with Pinnacle Actuarial Resources, Inc. in the San Francisco, California office. She holds a Master of Science degree in Statistics and a Bachelor of Science degree in Mathematics from Stanford University. Radost has over nine years of experience in the capacity of a pricing actuary in the personal lines segment. In this role, she has developed home and auto pricing solutions through the design and implementation of advanced predictive models.*